Research On The Deep Discriminative Feature Learning

Since 2012,deep learning has obtained remarkable successes and has shown its powerful modeling ability on data in computer vision.The modeling ability of deep model on data is instantiated in features,the discriminative feature has benefits for correct pattern classification,retrieval and matching,so as to meet the need of actual practice(in this thesis,features with good quality are called discriminative features).Therefore,how to learn and optimize the discrimina-tive deep feature is one of the challenging researches.Specifically,deep discriminative feature learning contains two problems,i.e.discrimination and generalization,which coexist and affect the discrimina-tive feature learning.However,deep discriminative feature learning is influ-enced by the objective loss function and model architectures.Thus,in order to ensure the discriminative feature learning,this paper needs to propose dif-ferent solutions for objective functions and model architectures.In summary,this thesis mainly focuses on the improvements of feature discrimination and generalization,the major contributions are as follows:1.For ensuring the learning of deep discriminative feature under Softmax,this thesis proposes Virtual-Softmax and Noisy-Softmax.Virtual-Softmax encourages the discrimination of features by a dynamic virtual class.Noisy-Softmax intends to use the annealing noise to improve the generalization of the learned features.These methods outperforms the conventional soft-max function and has been verified by extensive experiments on image classification and face recognition tasks.2.For ensuring the learning of deep discriminative feature under Deep-Metric-Learning,this thesis proposes two algorithms.The first is a graph-based auxiliary regularization term which can effectively help the current SOTA metric learning methods learn discriminative feature representations.The second is a confusion-based feature learning method,which is to improve the generalization of the learned feature.These two algorithms have been also demonstrated by popular image retrieval tasks.3.For ensuring the learning of deep discriminative feature by model archi-tecture,this thesis proposes mixed high-order attention network to im-prove the discrimination of features by combining the different outputs from different order.This method has been verified by the popular person re-identification task.